UCT CS Research Document Archive

How Evolvable in Novelty Search?

Shorten, David and Geoff Nitschke (2014) How Evolvable in Novelty Search? . In Proceedings IEEE Symposium Series in Computational Intelligence (SSCI 2014).

Full text available as:
PDF - Requires Adobe Acrobat Reader or other PDF viewer.

Abstract

This research compares the efficacy of novelty versus objective based search for producing evolvable populations in the maze solving task. Populations of maze solving simulated robot controllers were evolved to solve a variety of different, relatively easy, mazes. This evolution took place using either novelty or objective-based search. Once a solution was found, the simulation environment was changed to one of a variety of more complex mazes. Here the population was evolved to find a solution to the new maze, once again with either novelty or objective based search. It was found that, regardless of whether the search in the second maze was directed by novelty or fitness, populations that had been evolved under a fitness paradigm in the first maze were more likely to find a solution to the second. These results suggest that populations of controllers adapted under novelty search are less evolvable compared to objective based search in the maze solving task.

EPrint Type:Conference Paper
Subjects:I Computing Methodologies: I.2 ARTIFICIAL INTELLIGENCE
ID Code:974
Deposited By:Nitschke, Geoff
Deposited On:28 September 2014